Clémentin BOITTIAUX (Nov. 2020)
- Vincent HUGEL, University Professor (director),
- Claire DUNE, Associate Professor at the University of Toulon (co-supervisor),
- Aurélien ARNAUBEC, IFREMER researcher, PRAO team, IFREMER, la Seyne-sur-Mer (co-supervisor),
- Ricard MARXER, Lecturer at the University of Toulon (co-supervisor).
During the deployment of an underwater vehicle at high depths, absolute positioning techniques do not allow to reach an accuracy higher than 10m. Several techniques exist to improve navigation accuracy, but the residual positioning error in the mapping data must be corrected manually by the end user from the acoustic or optical images of the bottom. In the case of an autonomous submarine (AUV), the system must be able to (re)locate itself from the analysis of data collected by its sensors. In terrestrial robotics, localization algorithms are based on the detection of stable geometric primitives anchored on, for example, the salient angles of buildings. These algorithms are not very robust in unstructured and changing environments such as underwater environments. Therefore, other more robust descriptors must be considered that allow an AUV to constrain the drift of the proper positioning error of estimation-based navigation systems. The solution that is proposed in this research work is based on real-time and delayed-time information processing to allow an autonomous underwater robot to relocate in its environment.
The proposed research work will answer the following questions:
- To what extent can recent advances in artificial intelligence offer interesting solutions for robust semantic segmentation of underwater spaces?
During underwater campaigns, a large amount of data is acquired by both cameras and sonars. This data could be used to feed a deep learning system. The problem lies in the limited annotation of this data.
- Is it possible to train a deep learning system from data from multiple modalities with little annotation?
- Moreover, the acquisition system is an active system: the AUV’s movements, the modifications of its lighting, the choice of sensors can improve the detection. How to control the AUV in order to obtain optimal data for terrain recognition?
The navigation strategy can take into account a priori information (previous maps, specific observations, tracking of terrain features) to optimize the information obtained to locate the robot